Pith. sign in

REVIEW 2 major objections 7 minor 51 references

Industrial3D: A Water-Treatment TLS Point Cloud Dataset and Cross-Paradigm Benchmark for MEP Scene Understanding

T0 review · 2 major / 7 minor · reviewed 2026-07-13 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Industrial water-treatment TLS scans remain far harder than architectural ones: best supervised mIoU is 55.74%, zero-shot foundation models only 15.79%.

desk verdict Solid, useful industrial TLS dataset paper: real scale, honest cross-paradigm numbers, and a clear domain gap; scope is water-treatment only, but that does not undercut the measurement. read the letter →

arxiv 2603.28660 v2 pith:KQGKIKCP submitted 2026-03-30 cs.CV

Chao Yin , Hongzhe Yue , Qing Han , Difeng Hu , Zhenyu Liang , Fangzhou Lin , Bing Sun , Boyu Wang
show 3 more authors
Mingkai Li Wei Yao Jack C.P. Cheng
This is my paper · ORCID
classification cs.CV
keywords terrestriallaserscanningpointcloudsemanticsegmentationindustrialMEPScan-to-BIMclassimbalancefoundationmodelsweaklysupervisedlearningdigitaltwin
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that automatic semantic labeling of dense laser scans of real industrial plants is still largely unsolved, because those scenes are not like offices. Water-treatment facilities pack pipes, valves, flanges, and pumps into dense, occluded layouts with extreme class imbalance and shared cylindrical geometry, so models trained on architectural benchmarks fail to transfer. The authors release Industrial3D—612 million expert-labeled points at 6 mm resolution from 13 areas in operational plants, 6.6 times larger than the nearest MEP dataset—and run a unified benchmark of nine methods across fully supervised, weakly supervised, unsupervised, and foundation-model settings. The best supervised result is only 55.74% mean IoU; zero-shot Point-SAM reaches 15.79%, a 39.95-point gap the authors attribute to a dual crisis of 215:1 rarity and geometric ambiguity between rare fittings and ordinary pipes. A sympathetic reader cares because Scan-to-BIM, digital twins, and as-built checks for industrial infrastructure all depend on closing that gap.

What carries the argument

The dual crisis: simultaneous 215:1 head-to-tail statistical rarity and shared cylindrical primitives that make rare fittings (valves, flanges, reducers, strainers) locally indistinguishable from abundant pipes, so frequency re-weighting alone cannot separate them.

What would settle it

Train and evaluate the same nine methods on an equally large TLS corpus from a different industrial plant type (power, oil-and-gas, or manufacturing) with comparable class taxonomy; if the supervised–zero-shot gap collapses below ~20 points or tail-class IoUs rise without geometry-aware losses, the dual-crisis account and the claimed domain distance would be falsified.

Watch

Extended reading notes

Core claim

Industrial3D is the largest public industrial MEP TLS semantic-segmentation testbed (612.7M labeled points, 12 classes, 13 areas). Under one protocol the best supervised method reaches 55.74% mIoU while zero-shot Point-SAM reaches 15.79%—a 39.95-point gap the authors attribute to a dual crisis of 215:1 class imbalance and cylindrical geometric ambiguity between tail classes and head-class pipes.

Load-bearing premise

The claim that two held-out water-treatment areas and this plant type are representative enough of operational industrial MEP for the measured 40-point gap and dual-crisis diagnosis to generalize.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 7 minor

Summary. The paper introduces Industrial3D, a TLS point-cloud dataset of 612.7M expert-labeled points at 6 mm resolution from 13 areas / 7 operational water-treatment facilities, with 12 MEP/structural classes. It claims this is the largest public industrial MEP semantic-segmentation testbed (6.6× the closest comparable MEP set) and establishes a unified cross-paradigm benchmark of nine methods spanning fully supervised, weakly supervised, unsupervised, and foundation-model settings. Under that protocol the best supervised method (Boundary-CB) reaches 55.74% mIoU while zero-shot Point-SAM reaches 15.79%—a 39.95-point gap attributed to a dual crisis of 215:1 class imbalance and cylindrical geometric ambiguity between tail classes and head-class pipes. Supporting analyses include head/common/tail splits, SQN label-ratio ablations (including a PyTorch fidelity check on S3DIS), GrowSP with Hungarian matching, and Point-SAM oracle vs one-vs-rest protocols, with public release of data, code, and models promised.

Significance. If the measurements hold, Industrial3D fills a genuine gap: architectural RGB-D benchmarks (S3DIS, ScanNet) and smaller MEP sets do not capture the occlusion, cylindrical-primitive overlap, and extreme long-tail statistics of operational industrial plants. The scale enables previously infeasible 0.1% sparse-supervision and foundation-model transfer experiments, and the quantified 39.95-point gap plus dual-crisis diagnosis give the community a concrete, falsifiable target for Scan-to-BIM and digital-twin work. Strengths include the large expert-annotated corpus, area-based held-out split, multi-paradigm protocol under a single PyTorch stack, SQN reproduction fidelity on S3DIS (within ~1% of reported numbers), and planned public release of data, code, and weights—standard assets that make the empirical claims checkable.

major comments (2)
  1. Section 4.2 and Table 7: the headline supervised ceiling (55.74% mIoU) is achieved by Boundary-CB, which is cited as a companion methods paper (Yin et al., 2026) rather than a standard public baseline. For a dataset/benchmark paper this is acceptable only if the top result is reproducible from the promised release; otherwise the ceiling and the 39.95-point gap rest on an external method. Please either (i) ship full Boundary-CB training code/configs with the benchmark, or (ii) report the best fully public baseline (e.g., KPConv 53.65%) as the primary supervised number and treat Boundary-CB as an optional long-tail ablation.
  2. Section 4.1 / Table 6 and Discussion §6: the dual-crisis diagnosis and the 39.95-point gap are measured on a single area-based split (Areas 6+12 held out) within water-treatment plants only. The manuscript already notes limited domain coverage, but the abstract and contributions still frame the gap as characterizing “industrial TLS data” more broadly. A leave-one-area-out or multi-split sensitivity check (even on a subset of methods) would substantially strengthen the claim that the gap is not an artifact of this particular train/test assignment; without it, the generalization language should be tightened to “within-domain water-treatment MEP.”
minor comments (7)
  1. Abstract vs body: abstract says “20 room scenes, 13 dataset areas, and 7 operational water treatment facilities”; body sometimes says “13 water treatment facilities” (e.g., Abstract of the full text, §3). Align the facility/area/scene counts everywhere.
  2. Table 1 and §2.1: imbalance ratios for prior datasets are approximate or missing (“–”); state how 215:1 for Industrial3D is computed (Rbm:Str from Table 5) and whether prior ratios use the same definition.
  3. §4.3.1: IoU@k adaptation for Point-SAM (oracle vs one-vs-rest) is clear, but the exact sampling of the five positive/negative prompts and whether prompts are class-balanced should be stated so the 15.79% number is fully reproducible.
  4. §4.4.2 / Table 8: SQN is only paired with RandLA-Net; a short note that stronger backbones remain future work is already present—consider elevating it so readers do not over-generalize the “0.1% beats 100%” claim beyond that family.
  5. Figures 5–6 and Table 5: log-scale and mixed K/M units are dense; a single supplementary CSV of per-area per-class counts would aid re-use.
  6. Typos / consistency: “verifcation” (abstract), “quantifes”, occasional “Industrial3D” vs “Industrial 3D”, and mixed “Rbm/RectBeam/rbeam” abbreviations—standardize.
  7. No multi-seed error bars or variance on mIoU (reader note). Even 2–3 seeds for the main supervised and SQN rows would help; if compute-limited, state that explicitly.

Circularity Check

0 steps flagged · score 1.0 of 10

Empirical dataset/benchmark paper: measured mIoU gaps and class statistics do not reduce by construction to inputs; only mild non-load-bearing self-citation of companion Boundary-CB and prior PSNet5/ResPointNet++.

full rationale

Industrial3D is a dataset-and-benchmark paper, not a first-principles derivation. The load-bearing claims are empirical: (i) 612.7M labeled points and 215:1 head-to-tail ratio from Table 5 / Fig. 5; (ii) held-out Areas 6+12 mIoU for nine methods (Tables 7–10), including zero-shot Point-SAM at 15.79% vs best supervised 55.74% (39.95 pp gap); (iii) head–tail disparity under class-balanced losses (Table 11). None of these quantities is fitted then re-presented as a prediction, nor defined in terms of the result they are said to establish. Self-citations to PSNet5, ResPointNet++, Yin et al. 2023, and the companion Boundary-CB (Yin et al., 2026) supply a prior dataset, a baseline architecture, and one of six supervised methods; the gap and dual-crisis diagnosis remain intact if Boundary-CB is removed (KPConv 53.65%, ResPointNet++ 52.48% still leave a ~37–38 pp foundation-model gap and near-zero strainer/reducer IoU). Geometric-ambiguity language (tail fittings share cylindrical primitives with pipes) follows from a standard MEP taxonomy and observed confusions, not from a uniqueness theorem or ansatz smuggled in as external fact. Scope limits (water-treatment-only corpus; Areas 6/12 split) are stated in §6 and are external-validity concerns, not circularity. Score 1 reflects only routine author-overlap baselines, not a reduction of the central claim to its inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

This is an empirical dataset-and-benchmark paper. Load-bearing premises are domain and protocol choices (taxonomy, area split, metric adaptations, Hungarian matching for unsupervised, prompt protocols for Point-SAM) rather than free physical constants or invented particles. No mathematical free parameters are fitted to produce the headline gap; the gap is measured.

free parameters (4)
  • Sparse label ratios (0.01%, 0.1%) and SQN query/propagation settings
    Chosen following SQN practice; they define the weak-supervision operating points reported in Table 8.
  • Point-SAM prompt protocol (T=5 oracle iterations; 5 pos/5 neg one-vs-rest)
    Adaptation choices that convert instance-prompt IoU@k into class-level mIoU; they affect the reported foundation-model numbers.
  • Head/common/tail partition thresholds
    Classes grouped by frequency (head ~77%, common ~20%, tail <3% each) to define mIoU_head/tail and H-IoU; partition is design choice, not data-driven fit of a continuous parameter.
  • Train/test area assignment (Areas 6+12 held out)
    S3DIS-style area split chosen by authors; all headline mIoU numbers depend on this partition.
assumptions (5)
  • domain assumption Area-based train/test split without scene leakage is a valid generalization protocol for industrial TLS (S3DIS-style).
    Section 4.1 adopts entire-area holdout; validity depends on areas being exchangeable enough.
  • domain assumption Hungarian matching of unsupervised clusters to ground-truth classes yields a fair upper-bound mIoU for GrowSP.
    Section 4.3.1; standard but optimistic for unsupervised evaluation.
  • ad hoc to paper Adapted oracle and one-vs-rest prompt protocols make Point-SAM comparable to semantic mIoU baselines.
    Section 4.3.1 / 4.4.4; native Point-SAM metric is IoU@k on instances, not class mIoU.
  • domain assumption 12-class MEP taxonomy plus exclusion of floors/walls/unclassified is the right semantic target for industrial Scan-to-BIM evaluation.
    Section 3.2 Table 3; alternative taxonomies would change imbalance and scores.
  • domain assumption Water-treatment facilities are representative enough of operational industrial MEP for the dual-crisis claim.
    Introduction and §6; authors themselves note limited facility-type coverage.
invented entities (2)
  • Dual crisis (statistical rarity + cylindrical geometric ambiguity)
    purpose: Named diagnosis for why long-tail industrial MEP segmentation fails under reweighting and geometry-only learning.
    Useful framing of observed failure modes; not an independent physical entity, and evidence is internal to this benchmark.
  • Industrial3D dataset (612.7M labeled TLS points, 12 classes, 13 areas) independent evidence
    purpose: Provide a public industrial MEP testbed larger and more imbalanced than prior architectural/MEP sets.
    Core contribution; independent of any single model once released.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Industrial3D: A Water-Treatment TLS Point Cloud Dataset and Cross-Paradigm Benchmark for MEP Scene Understanding." pith.science (2026). https://pith.science/paper/KQGKIKCP

@misc{pith2026260328660,
  author       = {Pith},
  title        = {Pith review of: Industrial3D: A Water-Treatment TLS Point Cloud Dataset and Cross-Paradigm Benchmark for MEP Scene Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KQGKIKCP}},
  note         = {Machine review of arXiv:2603.28660}
}
read the original abstract

Automated semantic understanding of dense terrestrial laser scanning (TLS) point clouds is a prerequisite for Scan-to-BIM, digital twin maintenance, and as-built verifcation. Yet for operational industrial mechanical, electrical, and plumbing (MEP) facilities, this challenge remains largely unsolved: water-treatment TLS scans exhibit extreme geometric ambiguity, severe occlusion, and extreme class imbalance that architectural benchmarks such as S3DIS and ScanNet cannot adequately represent. We present Industrial3D, a terrestrial LiDAR dataset with 612.7 million expert-labeled points at 6 mm resolution from 20 room scenes, 13 dataset areas, and 7 operational water treatment facilities. At 6.6x the scale of the closest comparable MEP dataset, Industrial3D provides the largest industrial MEP testbed for within-domain scene understanding. We further establish a cross-paradigm benchmark of nine methods across fully supervised, weakly supervised, unsupervised, and foundation-model settings. The best supervised method reaches 55.74% mIoU, whereas zero-shot Point-SAM reaches 15.79%, a 39.95 percentage-point gap that quantifes unresolved domain transfer for industrial TLS data. Analysis attributes this gap to a dual crisis: 215:1 statistical rarity and cylindrical geometric ambiguity between tail classes and head-class pipes. The dataset, benchmark code, and pre-trained models will be publicly released at https://github.com/pointcloudyc/Industrial3D.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 17 canonical work pages

  1. [1]

    Advanced Engineering Informatics 45, 101121

    Cloi-net: Class segmentation of industrial facilities’ point cloud datasets. Advanced Engineering Informatics 45, 101121. doi:10.1016/j.aei.2020.101121. Agapaki, E., Glyn-Davies, A., Mandoki, S., Brilakis, I.,

    work page doi:10.1016/j.aei.2020.101121 2020

  2. [2]

    Computing in Civil Engineering 33, 66–73

    Cloi: A shape classification benchmark dataset for industrial facilities. Computing in Civil Engineering 33, 66–73. doi:10.1061/(ASCE)CP.1943-5487.0000843. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.,

    work page doi:10.1061/(asce)cp.1943-5487.0000843 1943

  3. [3]

    1534–1543

    3d semantic parsing of large-scale indoor spaces, in: Proceedings of the IEEE conference on computer vision and pattern recog- nition, pp. 1534–1543. doi:10.1109/CVPR.2016.170. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.,

    doi:10.1109/cvpr.2016.170 2016

  4. [4]

    doi:10.1109/ICCV.2019.00939

    Semantickitti: A dataset for semantic scene understand- ing of lidar sequences, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). doi:10.1109/ICCV.2019.00939. 40 Chen, Z., Xu, H., Chen, W., Zhou, Z., Xiao, H., Sun, B., Xie, X., Kang, W.,

    work page doi:10.1109/iccv.2019.00939 2019

  5. [5]

    14244–14253

    Pointdc: Unsupervised semantic segmentation of 3d point clouds via cross-modal distillation and super-voxel clustering, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 14244–14253. doi:10.1109/ICCV51070.2023.01314. CloudCompare,

    doi:10.1109/iccv51070.2023.01314 2023

  6. [6]

    5828–5839

    Scannet: Richly-annotated 3d reconstructions of indoor scenes, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 5828–5839. doi:10.1109/CVPR.2017.261. Gao, W., Peters, R., Stoter, J.,

    doi:10.1109/cvpr.2017.261 2017

  7. [7]

    ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences X-4/W5-2024, 179–186

    Building-pcc: Building point cloud completion benchmarks. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences X-4/W5-2024, 179–186. doi:10. 5194/isprs-annals-X-4-W5-2024-179-2024,arXiv:2404.15644. Graham, B., van der Maaten, L.,

    arXiv 2024

  8. [9]

    IEEE transactions on pattern anal- ysis and machine intelligence 43, 4338–4364

    Deep learning for 3d point clouds: A survey. IEEE transactions on pattern anal- ysis and machine intelligence 43, 4338–4364. doi:10.1109/TPAMI.2020. 3005434. Hackel, T., Savinov, N., Ladicky, L., Wegner, J.D., Schindler, K., Polle- feys, M.,

    doi:10.1109/tpami.2020 2020

Show all 51 references
  1. [10]

    net: A new large-scale point cloud clas- sification benchmark

    Semantic3d. net: A new large-scale point cloud clas- sification benchmark. arXiv preprint arXiv:1704.03847 doi:10.5194/ isprs-annals-iv-1-w1-91-2017. Hu, D., Gan, V.J., Yin, C.,

    2017 arXiv

  2. [11]

    Automation in Construction 152, 104949

    Robot-assisted mobile scanning for auto- mated 3d reconstruction and point cloud semantic segmentation of build- ing interiors. Automation in Construction 152, 104949. doi:10.1016/j. autcon.2023.104949. Hu, Q., Yang, B., Fang, G.C., Guo, Y., Leonardis, A., Trigoni, N., Markham, A.,

    2023 doi

  3. [12]

    Sqn: Weakly-supervised semantic segmentation of large-scale 41 3d point clouds, in: Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXVII, Springer. pp. 600–619. doi:10.1007/978-3-031-19827-4_34, arXiv:2104.04891. H...

    2022 doi arXiv

  4. [13]

    11108–11117

    Randla-net: Efficient semantic segmentation of large-scale point clouds, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11108–11117. doi:10.1109/CVPR42600.2020. 01112. Huang, K.C., Li, X., Qi, L., Yan, S., Yang, M.H.,

    2020 doi

  5. [14]

    Reason3d: Searching and reasoning 3d segmentation via large language model, in: International Conference on 3D Vision (3DV), pp. 1–12. doi:10.48550/arXiv.2405. 17427. Huang, R., Xu, Y., Hong, D., Yao, W., Ghamisi, P., Stilla, U.,

    doi

  6. [15]

    ISPRS Journal of Photogrammetry and Re- mote Sensing 163, 62–81

    Deep point embedding for urban classification using als point clouds: A new per- spective from local to global. ISPRS Journal of Photogrammetry and Re- mote Sensing 163, 62–81. doi:https://doi.org/10.1016/j.isprsjprs. 2020.02.020. Jing, S., Zhong, R., Li, X., Aung, P.P.W., Par...

    2020 doi

  7. [16]

    Journal of Building Engineering 96, 110311

    Improved building mep systems semantic segmentation in point clouds using a novel multi-class dataset and local–global vector transformer network. Journal of Building Engineering 96, 110311. doi:10.1016/j.jobe.2024.110311. Khanzode, A., Fischer, M., Reed, D.,

    2024 doi

  8. [17]

    Journal of Information Technology in Construc- tion (ITcon) 13, 324–342

    Benefits and lessons learned of implementing building virtual design and construction (vdc) technologies for coordination of mechanical, electrical, and plumbing (mep) systems on a large healthcare project. Journal of Information Technology in Construc- tion (ITcon) 13, 324–34...

    2008 doi

  9. [18]

    2684–2693

    Large-scale point cloud semantic seg- mentation with superpoint graphs, in: Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pp. 2684–2693. doi:10.1109/CVPR.2018.00286,arXiv:1711.09869. Li, M., Gan, V.J., Wang, B.,

    2018 doi

  10. [19]

    Automation in Construction 179, 106455

    Integrating hierarchical segmentation and vision-language reasoning for spatially complex and occluded mep point 42 clouds. Automation in Construction 179, 106455. doi:10.1016/j.autcon. 2025.106455. Lin, X., Lin, T., Huang, L., Xie, H., Su, Z.,

    2025 doi

  11. [20]

    Liu, L., Zhuang, Z., Huang, S., Xiao, X., Xiang, T., Chen, C., Wang, J., Tan, M.,

    Bip3d: Bridging 2d images and 3d perception for embodied intelligence, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).arXiv:2411.14869. Liu, L., Zhuang, Z., Huang, S., Xiao, X., Xiang, T., Chen, C., Wang, J., Tan, M.,

    arXiv

  12. [21]

    15621–15632

    CPCM: Contextual point cloud modeling for weakly-supervised point cloud semantic segmentation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 15621–15632. doi:10. 48550/arXiv.2307.10316,arXiv:2307.10316. Liu, Z., Hu, H., Cao, Y., Zhang, Z., Tong, X.,

    arXiv

  13. [22]

    A closer look at local aggregation operators in point cloud analysis, in: European Conference on Computer Vision, pp. 358–374. doi:10.1007/978-3-030-58592-1_20. Liu, Z., Qi, X., Fu, C.W.,

    doi

  14. [23]

    1726–1736

    One thing one click: A self-training approach for weakly supervised 3d semantic segmentation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 1726–1736. doi:10.1109/CVPR46437.2021.00177. Pierdicca, R., Paolanti, M., Matrone, F., Marti...

    2021 doi

  15. [24]

    Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.,

    doi:10.3390/rs12061005. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.,

    doi

  16. [25]

    5648–5656

    Volumet- ric and multi-view cnns for object classification on 3d data, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 5648–5656. doi:10.1109/CVPR.2016.609. Qi, C.R., Yi, L., Su, H., Guibas, L.J.,

    2016 doi

  17. [26]

    5099–5108

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space, in: Advances in neural in- formation processing systems, pp. 5099–5108. doi:10.48550/arXiv.1706. 02413. Rauch, L., Braml, T.,

    doi

  18. [27]

    IEEE Transactions on Geoscience and Remote Sensing 62, 1–12

    Urban geobim con- struction by integrating semantic lidar point clouds with as-designed bim models. IEEE Transactions on Geoscience and Remote Sensing 62, 1–12. doi:10.1109/TGRS.2024.3358370. Tang, P., Huber, D., Akinci, B., Lipman, R., Lytle, A.,

    2024 doi

  19. [28]

    Automation in Construction 19, 829–843

    Automatic reconstruction of as-built building information models from laser-scanned point clouds: A review of related techniques. Automation in Construction 19, 829–843. doi:10.1016/j.autcon.2010.06.007. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.,

    2010 doi

  20. [29]

    6411–6420

    Kpconv: Flexible and deformable convolution for point clouds, in: ProceedingsoftheIEEE/CVFinternationalconferenceoncomputervi- sion, pp. 6411–6420. doi:10.1109/ICCV.2019.00651,arXiv:1904.08889. Wang, B., Wang, Q., Cheng, J.C., Song, C., Yin, C.,

    2019 doi

  21. [30]

    Automation in Construction 133, 103997

    Vision-assisted bim reconstruction from 3d lidar point clouds for mep scenes. Automation in Construction 133, 103997. doi:10.1016/j.autcon.2021.103997. Wang, H., Liu, Q., Yue, X., Lasenby, J., Kusner, M.J.,

    2021 doi

  22. [31]

    9782–9792

    Unsupervised point cloud pre-training via occlusion completion, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9782–9792. doi:10.1109/ICCV48922.2021.00972,arXiv:2010.01089. Wang, H., Rong, X., Yang, L., Wang, S., Tian, Y.,

    2021 doi

  23. [32]

    ISPRS Journal of Photogrammetry and Re- mote Sensing 188, 237–254

    A new weakly supervised approach for als point cloud semantic segmentation. ISPRS Journal of Photogrammetry and Re- mote Sensing 188, 237–254. doi:10.1016/j.isprsjprs.2022.04.016. Wang, P., Yao, W., Shao, J.,

    2022 doi

  24. [33]

    ISPRS Journal of Photogrammetry and Remote Sensing 204, 89–104

    One class one click: Quasi scene-level weakly supervised point cloud semantic segmentation with active learn- ing. ISPRS Journal of Photogrammetry and Remote Sensing 204, 89–104. doi:10.1016/j.isprsjprs.2023.09.002. Wang, T., Mao, X., Zhu, C., Xu, R., Lyu, R., Li, P., Chen, X....

    2023 doi

  25. [34]

    doi:10.1109/CVPR52733.2024.01868

    Embodiedscan: 44 A holistic multi-modal 3d perception suite towards embodied ai, in: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). doi:10.1109/CVPR52733.2024.01868. Wu, X., DeTone, D., Frost, D., Shen, T., Xie, C., Yang, N., Engel,...

    2024 doi

  26. [35]

    22193–22204

    Sonata: Self-supervised learning of reliable point representations, in: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pp. 22193–22204. doi:10.1109/CVPR52734.2025.02067,arXiv:2503.16429. Wu, X., Jiang, L., Wang, P.S., Liu, Z., Liu, X., Qia...

    2025 doi

  27. [36]

    4840–4851

    Point transformer v3: Simpler, faster, stronger, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pp. 4840–4851. doi:10.1109/CVPR52733.2024.00463, arXiv:2312.10035. Xiao, A., Zhang, X., Shao, L., Lu, S.,

    2024 doi

  28. [37]

    IEEE Transactions on Pattern Analysis and Machine Intelligence doi:10.1109/TPAMI.2024.3416302

    A survey of label-efficient deep learning for 3d point clouds. IEEE Transactions on Pattern Analysis and Machine Intelligence doi:10.1109/TPAMI.2024.3416302. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L.J., Litany, O.,

    2024 doi

  29. [38]

    Point- contrast: Unsupervised pre-training for 3d point cloud understanding, in: European Conference on Computer Vision, pp. 574–591. doi:10.1007/ 978-3-030-58580-8_34,arXiv:2007.10985. Xu, X., Lee, G.H.,

    2007 arXiv

  30. [39]

    13706–13715

    Weakly supervised semantic point cloud segmen- tation: Towards 10x fewer labels, in: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pp. 13706–13715. doi:10.1109/CVPR42600.2020.01368. Yeo, C., Kim, S., Kim, H., Kim, S., Mun, D.,

    2020 doi

  31. [40]

    JMST Advances 2, 15–24

    Deep learning ap- plications in an industrial process plant: repository of segmented point clouds for pipework components. JMST Advances 2, 15–24. doi:10.36827/ jmst-adv.2020.00003. Yin, C., Cheng, J.C., Wang, B., Gan, V.J.,

    2020

  32. [41]

    Automation in Construction 139, 104300

    Automated classifica- tion of piping components from 3d lidar point clouds using se-pseudogrid. Automation in Construction 139, 104300. doi:10.1016/j.autcon.2022. 104300. 45 Yin, C., Han, Q., Hou, Z., Liu, Y., Dai, A., Hu, H., Yang, J., Yao, W.,

    2022 doi

  33. [42]

    arXiv preprint arXiv:2601.19128arXiv:2601.19128

    Resolving primitive-sharing ambiguity in long-tailed industrial point cloud segmentation via spatial context constraints. arXiv preprint arXiv:2601.19128arXiv:2601.19128. Yin, C., Wang, B., Cheng, J.,

    arXiv

  34. [43]

    Automation in Construction 130, 103874

    Automated semantic segmentation of industrial point clouds using respointnet++. Automation in Construction 130, 103874. doi:10.1016/j.autcon.2021. 103874. Yin, C., Yang, B., Cheng, J.C., Gan, V.J., Wang, B., Yang, J.,

    2021 doi

  35. [44]

    Automation in Construction 148, 104757

    Label- efficient semantic segmentation of large-scale industrial point clouds us- ing weakly supervised learning. Automation in Construction 148, 104757. doi:10.1016/j.autcon.2023.104757. Yue, H., Wang, Q., Nie, X., Fang, H., Cheng, J.C.P., Jing, S., Wang, B.,

    2023 doi

  36. [45]

    Advanced Engineering Informatics 72, 104453

    Point cloud instance segmentation for building indoor scenes using deep learning and bim-generated synthetic point clouds. Advanced Engineering Informatics 72, 104453. doi:10.1016/j.aei.2026.104453. Yue, H., Wang, Q., Zhao, H., Zeng, N., Tan, Y.,

    2026 doi

  37. [46]

    Automation in Construction 168, 105769

    Deep learning ap- plications for point clouds in the construction industry. Automation in Construction 168, 105769. doi:10.1016/j.autcon.2024.105769. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.,

    2024 doi

  38. [47]

    8552–8562

    PointCLIP: Point cloud understanding by CLIP, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pp. 8552–8562. doi:10.1109/CVPR52688.2022.00836, arXiv:2112.02413. Zhang, Y., Qu, Y., Xie, Y., Li, Z., Zheng, S., Li, C.,

    2022 doi

  39. [48]

    doi:10.1109/ICCV48922.2021.01708

    Perturbed self-distillation: Weakly supervised large-scale point cloud semantic seg- mentation, in: Proceedings of the IEEE/CVF International Conference on ComputerVision, pp.15500–15508. doi:10.1109/ICCV48922.2021.01708. 46 Zhang, Y., Wu, X., Lao, Y., Wang, C., Tian, Z., Wang...

    2021 doi

  40. [49]

    Concerto: Joint 2d-3d self-supervised learning emerges spatial representa- tions, in: Advances in Neural Information Processing Systems, pp. 1–17. arXiv:2510.23607. Zhang, Z., Yang, B., Wang, B., Li, B.,

  41. [50]

    17619–17629

    Growsp: Unsupervised semantic segmentation of 3d point clouds, in: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pp. 17619–17629. doi:10.1109/CVPR52729.2023.01690,arXiv:2305.16404. Zhou, Y., Gu, J., Chiang, T.Y., Xiang, F., Su, H.,

    2023 doi

  42. [51]

    Point-SAM: Promptable 3d segmentation model for point clouds, in: The Thir- teenth International Conference on Learning Representations, pp. 1–18. doi:10.48550/arXiv.2406.17741,arXiv:2406.17741. Zhou, Z.H.,

    doi arXiv

  43. [52]

    arXiv preprint arXiv:2201.07882 doi:10.48550/arXiv.2201.07882

    A comprehensive survey on weakly supervised learning. arXiv preprint arXiv:2201.07882 doi:10.48550/arXiv.2201.07882. 47

    doi arXiv

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.